On March 29, 2026, Meta president Dina Powell McCormick took the stage at the Axios AI Summit in Washington and said something that didn’t sound like a typical AI keynote: the US needs 500,000 new electricians within two years to build what the AI industry has already promised.

Four months later, on July 21, Ford, Google, BlackRock, and Carhartt confirmed they believed her. They put $450 million combined behind the problem.

The bottleneck on the next phase of AI isn’t chips. It’s hands

For two years, every conversation about AI’s limits circled the same thing: how many GPUs you can get, how much compute costs, how fast you can train the next model. That framing has a problem that’s only now becoming visible: you can buy more compute capacity with enough capital, almost immediately. You cannot buy a certified electrician the same way.

Training an electrician who can safely wire a substation or a next-generation data center takes years of regulated apprenticeship, certification exams, and supervised field hours. That process doesn’t speed up by throwing more money at it, the same way a pregnancy doesn’t speed up by paying nine different women more. It’s a structural bottleneck, not one capital solves on its own, and it’s exactly the kind of constraint the public AI conversation has almost entirely ignored.

The numbers behind the warning

Per MachineBrief, McCormick was direct about the scale of the problem: the US will need 500,000 new electricians in two years to sustain the physical infrastructure behind AI’s expansion, from data centers to substations to charging networks.

The corporate response came fast. Per a report breaking down the commitments, Ford put in $300 million as part of its “Essential Economy” initiative, BlackRock committed $100 million to connect 50,000 workers with training over five years, and Google earmarked $50 million to prepare more than 300,000 workers across 20-plus states, partnered with 14 unions and four trade and contractor associations. Carhartt added its existing “For the Love of Labor” program. Together, the four companies launched the “Alliance for America’s Skilled Trades” on July 21, 2026.

Ford CEO Jim Farley put the problem in concrete terms for his own industry, per Allwork.Space: “by 2029, we’re going to need more than 350,000 new auto technicians across the country.” He added the part I find most interesting about the whole initiative: “no one person or organization is going to solve the skilled trades shortage alone.”

The scale of the problem isn’t unique to AI. The construction sector had 298,000 open positions as of May 2026, and the Manufacturing Institute projects 2.1 million unfilled skilled-trades positions by 2030. What changed is that demand for electricians, HVAC technicians, and linemen to wire AI data centers suddenly stacked on top of a shortage that was already building before any of this.

Why this doesn’t get solved with more capital

We already wrote about why AI is infrastructure, not a tool, and this is the most literal example of that idea I’ve seen yet. A tool gets installed. Infrastructure gets built, connected to a real power grid, and needs qualified people to keep it running for decades. No amount of additional money shortens how long it takes to certify an electrician, the same way no amount of money shortens how long it takes a language model to learn basic physics it’s never seen before.

This connects to something we already documented in the installer economy Silicon Valley just rediscovered: software scales by copying code. Skilled physical labor doesn’t scale by copying anything; it scales by training people, one at a time, with years in between. And as we already saw in the Jevons paradox and the jobs that doubled when the factory went dark, automation doesn’t always reduce demand for human labor. Sometimes it just shifts it to a different kind of work, one that turns out to be far harder to scale quickly than the one it replaced.

What this means for any company building on AI

If your company depends on AI infrastructure, whether you’re building data centers, expanding compute capacity, or simply planning what the next generation of models will cost and when it’ll be ready, this announcement leaves a concrete lesson: physical infrastructure timelines, not just model release announcements, should be part of any large-scale AI plan.

If a roadmap assumes power, connectivity, and maintenance will be ready exactly when the model is, it ignores the real bottleneck Meta, Ford, and Google have already flagged in public, with numbers and $450 million behind it. At IQ Source, when we help a company plan AI adoption at scale, the discovery question isn’t just how capable the model you’re going to use is. It’s whether the physical infrastructure behind that model, and the qualified people to maintain it, will be ready when you need it, or whether you’ll discover the shortage once it’s too late to solve with budget alone.

Plan your AI infrastructure before the bottleneck becomes yours